Task cards

bench/registry/tasks.yaml holds one card per task: dataset and split rule, input window, output and decoder, optimisation recipe per protocol, selection rule, metrics with their direction, scorer. imu-bench tasks prints the summary. The ten tasks of the benchmark:

Task id	Dataset	Input window	Output	Primary metric	Split
arm_pose_nymeria	Nymeria, real right-wrist IMU + mocap	10.24 s, 50 Hz	5 right-arm joint rotations per 0.32 s	position error, cm ↓	recording session
full_body_pose_nymeria	Nymeria	10.24 s	23 SMPL joint rotations per 0.32 s	root-relative MPJPE, mm ↓	recording session
odometry_nymeria_local	Nymeria	2.56 s	pelvis velocity in the wrist frame, integrated	ATE, m ↓	recording session
microgesture_openwatch	OpenWatch (43 canonical identities)	1 s, 100 Hz	5 gestures + background, voted per clip	clip macro-F1 ↑	subject, posture-stratified
language_humanml3d	HumanML3D, simulated wrist IMU	10.24 s	alignment with CLIP ViT-B/32 text	mean Recall@1, galleries of 100 ↑	official test split
pd_pads	PADS (469 participants)	2.56 s	healthy / Parkinson's / differential per participant	participant macro-F1 ↑	subject-disjoint 80/20
har_hhar	HHAR (6 classes)	2.56 s	window label	macro-F1 ↑	subject-disjoint
har_mhealth	MHEALTH (12)	2.56 s	window label	macro-F1 ↑	subject-disjoint
har_prism	PrISM (5 global classes)	2.56 s	window label	macro-F1 ↑	subject-disjoint
har_userwatch	TransfHAR (60)	2.56 s	window label	macro-F1 ↑	80/20 in time per recording

Every input is one wrist, accelerometer x/y/z (m/s², gravity included) then gyroscope x/y/z (rad/s), in a forward–left–up frame, resampled to the card's rate.

Decoders: linear heads for classification; width-192 attention decoders with two cross-channel and two temporal layers for arm pose, one cross-channel layer for full-body pose and one temporal layer for odometry; a projection into the CLIP text space for retrieval. Classification heads are trained with inverse-frequency class-weighted cross-entropy. The activity column of the leaderboard is the mean macro-F1 over the four activity sets.

Other task ids in tasks.yaml (decoder ladders, rotation variants, the release split of OpenWatch, world-frame odometry, extra activity sets) are ablations and controls from the paper; they are marked disabled and need --allow-disabled in bench.launch.